An IoT-Based Automatic Product Weight Quality Control System Using Load Cell, HX711, Moving Average, and Fully Adaptive Kalman Filter

Abstract

Automated weight-based quality control (QC) is a critical requirement in food and beverage manufacturing, yet many production lines still rely on manual weighing, which is slow and prone to human error, while raw load-cell signals remain sensitive to mechanical vibration and electrical noise. This study designs and develops an Internet of Things (IoT)-based automatic weight QC system built around a load cell and an HX711 amplifier interfaced with an ESP32 microcontroller, an infrared sensor for product detection, and a servo motor for automatic PASS/REJECT sorting. The raw weight reading is processed sequentially through a Moving Average filter to suppress high-frequency fluctuation, followed by a Fully Adaptive Kalman Filter that continuously re-estimates its process- and measurement-noise covariances to track dynamic disturbances; the filtered weight is then compared against a 23 g target with a ±1 g tolerance, and the result is transmitted in real time to the Blynk application and logged to a Google Spreadsheet. Across 40 repeated weighing trials against a reference actual weight of 23.2 g, the system classified 33 products (82.5%) as PASS and 7 products (17.5%) as REJECT. The Fully Adaptive Kalman Filter produced a mean output of 22.785 g, compared with 22.882 g for the raw HX711 signal and 22.807 g for the Moving-Average-only output; overall accuracy remained close across all three stages (96.4-96.6%), with the raw and Moving Average stages showing marginally lower aggregate error than the Fully Adaptive Kalman stage under this repeated static-weight protocol, while the Kalman stage showed greater responsiveness to abrupt individual readings. These findings show that a cascaded Moving Average and Fully Adaptive Kalman Filter pipeline, combined with dual-channel IoT monitoring, yields a stable, traceable weight-QC platform suitable for deployment in industrial environments, while also indicating that the adaptive filter’s forgetting factor merits further tuning for near-constant setpoints.